Publications

As health data volume and the sophistication of artificial intelligence (AI) tools grow, their potential to transform the management of complex neuroimmunological conditions like multiple sclerosis (MS) has become increasingly evident (1). MS, a chronic immune-mediated inflammatory disorder of the central nervous system, presents a unique challenge in the health sector due to its multifactorial nature and variable progression patterns. Each patient's journey is marked by distinct symptom trajectories and responses to treatment, demanding personalised approaches in diagnosis, prognosis, and therapeutic interventions. (2, 3). This Special Topic aims to address the clinical complexity of MS by leveraging data driven insights and innovative health initiatives. The overarching goal is to present the current challenges in MS research and explore recent advances and future trends that can significantly impact patient care. Through a Research Topic of reviews, perspectives, and original research articles, we explore how advanced data techniques and innovative health initiatives are shaping the future of MS research and care.

The fragmented nature of Real-World Data (RWD) presents significant challenges for research, particularly in low-prevalence diseases like Multiple Sclerosis (MS). Data dispersed across institutions and countries, coupled with restrictive privacy regulations, limits access to comprehensive datasets, delaying the development of predictive models and understanding of critical outcomes. Federated Learning (FL) offers a collaborative solution by enabling model training without centralizing sensitive patient data, but it often fails to address variations across data providers. Personalized Federated Learning (PFL) adapts models to these differences, improving the precision and relevance of insights. Yet, its potential to enhance predictive performance and guide clinical decisions in MS using routine clinical RWD remains largely unexplored. In this study, we evaluate standard FL alongside two personalization strategies for predicting confirmed MS disability progression over two years, using RWD from over 26,000 patients in the MSBase registry. These strategies include (1) AdaptiveDualBranchNet, a novel architecture that selectively exchanges key model parameters, enabling more nuanced adaptation across diverse clinical centers and (2) fine-tuning the global FL model to better fit each local's data. We benchmark these methods against fully centralized (pooled data) and local (site-specific) models to assess their relative performance and practicality. Our results show that standard FL methods trail both centralized and local baselines, highlighting the limitations of a uniform global model. In contrast, personalization yields a notable performance boost, with both PFL strategies surpassing all comparators. Specifically, the adaptive and fine-tuned versions of FedProx and FedAVG achieved the highest ROC-AUC scores, with FedProx reaching 0.8398 ± 0.0019 and 0.8375 ± 0.0019, and FedAVG scoring 0.8384 ± 0.0014 and 0.8370 ± 0.0016, respectively. In contrast, among the standard FL methods, FedAdam and FedYogi showed the best performance with ROC-AUC scores of 0.7919 ± 0.0031 and 0.7910 ± 0.0028. These results underscore that personalization is not a luxury but a necessity, enabling FL to reach, and even exceed, the predictive power of traditional modeling approaches. By providing concrete guidelines for implementing PFL and introducing a new, flexible architecture, this study establishes a clear path toward unlocking FL's untapped potential. Ultimately, these advances establish a clear path for more impactful, privacy-aware predictive modeling that enhances clinical decision-making and patient care in MS and beyond.

Objective A novel method is introduced to regress over the Sign Language finger movements from human electrocorticography (ECoG) recordings. Approach The proposed Graph-Optimized Block-Term Tensor Regression (Go-BTTR) method consists of two components: a deflation-based regression model that sequentially Tucker-decomposes multiway ECoG data into a series of blocks, and a Causal Graph Process (CGP) that accounts for the complex relationship between finger movements when expressing sign language gestures. Prior to each regression block, CGP is applied to decide which fingers should be kept separate or grouped and should therefore be referred to BTTR or its extended version eBTTR, respectively. Main results Two ECoG datasets were used, one recorded in 5 patients expressing 4 hand gestures of the American Sign Language Alphabet, and another in 2 patients expressing all gestures of the Flemish Sign Language Alphabet. As Go-BTTR combines fingers in a flexible way, it can better account for the nonlinear relationship ECoG exhibits when expressing hand gestures, including unintentional finger co-activations. This is reflected by the superior joint finger trajectory predictions compared to eBTTR, and predictions that are on par with BTTR in single finger scenarios. Subject 5, Dataset 1, during the movement of hand gestures reached 0.82 Pearson correlation for the pinky for Go-BTTR compared to 0.8 and 0.72 for eBTTR and BTTR respectively. Significance Our findings show that Go-BTTR is capable of decoding complex hand gestures taken from the Sign Language Alphabet. Go-BTTR also demonstrates computational efficiency, providing a notable benefit when intracranial electrodes are inserted during a patient's pre-surgical evaluation. This efficiency helps reduce the time required for developing and testing a Brain-Computer Interface (BCI) solution.

Objective. To identify the electrocorticography (ECoG) frequency features that encode distinct finger movement states during repeated finger flexions. Approach. We used the publicly available Stanford ECoG dataset of cue-based, repeated single finger flexions. Using linear regression, we identified the spectral features that contributed most to the encoding of movement dynamics and discriminating movement events from rest, and combined them to predict finger movement trajectories. Furthermore, we also looked into the effect of the used frequency range and the spatial distribution of the identified features. Main results. Two frequency features generate superior performance, each one for a different movement aspect: high gamma band activity distinguishes movement events from rest, whereas the local motor potential (LMP) codes for movement dynamics. Combining these two features in a finger movement decoder outperformed comparable prior work where the entire spectrum was used as the average correlation coefficient with the true trajectories increased from 0.45 to 0.5, both applied to the Stanford dataset, and erroneous predictions during rest were demoted. In addition, for the first time, our results show the influence of the upper cut-off frequency used to extract LMP, yielding a higher performance when this range is adjusted to the finger movement rate. Significance. This study shows the benefit of a detailed feature analysis prior to designing the finger movement decoder.

Multiway- or tensor-based decoding techniques for Brain-Computer Interfaces (BCI) are believed to better account for the multilinear structure of brain signals than conventional vector- or matrix-based ones. However, despite their outlook on significant performance gains, the used parameter optimization approach is often too computationally demanding, so conventional techniques are still preferred. We propose two novel tensor factorizations which we integrate into our Block-Term Tensor Regression (BTTR) algorithm and further introduce a marginalization procedure that guarantees robust predictions while reducing the risk of overfitting (generalized regression). BTTR accounts for the underlying (hidden) data structure in a fully automatic and computationally-efficient manner, leading to a significant performance gain over conventional vector- or matrix-based techniques in a challenging real-world application. As a challenging real-world application, we apply BTTR to accurately predict single finger movement trajectories from intracranial recordings in human subjects. We compare the obtained performance with that of the state-of-the-art.

Objective We introduce extended Block-Term Tensor Regression (eBTTR), a novel regression method designed to account for the multilinear nature of human intracranial finger movement recordings. Approach The proposed method relies on recursive Tucker decomposition combined with automatic component extraction. Main results eBTTR outperforms state-of-the-art regression approaches, including multilinear and deep learning ones, in accurately predicting finger trajectories as well as unintentional finger coactivations. Significance eBTTR rivals state-of-the-art approaches while being less computationally expensive which is an advantage when intracranial electrodes are implanted acutely, as part of the patient's presurgical workup, limiting time for decoder development and testing.

Objective We introduce Sparse eLORETA, a novel method for estimating a nonparametric solution to the source localization problem. Its goal is to generate a sparser solution compared to other source localization methods including eLORETA while benefitting from the latter's superior source localization accuracy. Approach Sparse eLORETA starts by reducing the source space of the Lead Field Matrix using Structured Sparse Bayesian Learning (SSBL) from which a Reduced Lead Field Matrix is constructed, which is used as input to eLORETA. Main results With Sparse eLORETA, source sparsity can be traded against signal fidelity; the proposed optimum is shown to yield a much sparser solution than eLORETA's with only a slight loss in signal fidelity. Significance When pursuing a data-driven approach, for cases where it is difficult to choose specific regions of interest (ROIs), or when subsequently a connectivity analysis is performed, source space reduction could prove beneficial.

Directed connectivity between brain sources identified from scalp electroencephalography (EEG) can shed light on the brain's information flows and provide a biomarker of neurological disorders. However, as volume conductance results in scalp activity being a mix of activities originating from multiple sources, the correct interpretation of their connectivity is a formidable challenge despite source localization being applied with some success. Traditional connectivity approaches rely on statistical assumptions that usually do not hold for EEG, calling for a model-free approach. We investigated several types of Artificial Neural Networks in estimating Directed Connectivity between Reconstructed EEG Sources and assessed their accuracy with respect to several ground truths. We show that a Long Short-Term Memory neural network with Non-Uniform Embedding yields the most promising results due to its relative robustness to differing dipole locations. We conclude that certain network architectures can compete with the already established methods for brain connectivity analysis.

Objective: A case study is conducted to show that not only finger flexion- but also finger abduction trajectories can be decoded from high-density electrocorticography (ECoG) recordings. Approach: Two patients, temporarily implanted with high-density ECoG grids as part of their clinical workup, performed a cued single finger flexion task, a cued sign language alphabet task and a cued finger abduction task. The extended Block-Term Tensor Regression (eBTTR) model was adopted to predict finger trajectories after being developed on simultaneous data glove/ECoG recordings. The model performance on the single finger flexion task provides a reference for comparing that on the finger abduction task. Finally, the sign language alphabet task involves realistic, complex finger movements, calling upon both finger flexions and abductions. Main results: We show that finger abduction trajectories can be decoded from ECoG recordings, with comparable precision as single finger flexion trajectories, and even when expressing sign language gestures that involve joint finger flexions and abductions. Significance: Our findings show for the first time that finger abduction trajectories can be decoded from ECoG even when part of Sign Language Alphabet gestures.

Block-Term Tensor Regression (BTTR) is a multilinear modeling framework suited for high-dimensional biomedical data, but its centralized implementation conflicts with privacy and regulatory constraints. To overcome this, we propose Federated Block-Term Tensor Regression (FBTTR), which integrates BTTR into a federated learning setting. We systematically evaluate multiple federated aggregation strategies, including FedAvg, FedYogi, and FedAdam, to assess their effect on model performance and stability. Experiments on the BCI Competition IV dataset demonstrate that the choice of strategy strongly influences predictive accuracy: FedAvg yields the most stable and accurate results across subjects, while adaptive methods such as FedYogi and FedAdam show less consistent performance. These findings highlight the importance of strategy selection in federated learning and establish FBTTR as a practical approach for privacy-preserving analysis of multi-institutional biomedical data.

Block-Term Tensor Regression (BTTR) has proven to be a powerful tool for modeling complex, high-dimensional data by leveraging multilinear relationships, making it particularly well-suited for applications in healthcare and neuroscience. However, traditional implementations of BTTR rely on centralized datasets, which pose significant privacy risks and hinder collaboration across institutions. To address these challenges, we introduce Federated Block-Term Tensor Regression (FBTTR), an extension of BTTR designed for federated learning scenarios. FBTTR enables decentralized data analysis, allowing institutions to collaboratively build predictive models while preserving data privacy and complying with regulations. FBTTR represents a major step forward in applying tensor regression to federated learning environments. Its performance is evaluated in a case study: heart disease prediction. In this case study, FBTTR is applied to predict heart disease using real-world clinical datasets, outperforming both standard federated learning approaches and centralized BTTR models. In the Fed-Heart-Disease Dataset, an AUC-ROC was obtained of 0.872 ± 0.02 and an accuracy of 0.772 ± 0.02 compared to 0.812 ± 0.003 and 0.753 ± 0.007 for the centralized model. Empirical evaluations highlight the robustness and scalability of FBTTR, achieving predictive performance comparable to centralized models while significantly reducing privacy concerns. Additionally, the algorithm is computationally efficient, making it viable for real-world deployment where timely predictions are critical. By offering a scalable and privacy-preserving solution, FBTTR has the potential to advance predictive analytics across multiple domains, including healthcare and brain-computer interface applications. Notably, FBTTR is developed as open-source software, promoting transparency and fostering collaboration within the research community.

Controlling a hand exoskeleton based on electroencephalogram (EEG)-based brain-computer interfacing (BCI) holds promise for human motor augmentation and neurorehabilitation. To achieve natural control, a critical step is to understand the impact of proprioception provided by the exoskeleton during interaction. In this study, we aim to approach this goal by quantifying EEG representations and BCI performance. We monitored full-scalp EEG in 25 healthy subjects while they performed different finger movement tasks with a cable-driven hand exoskeleton. Each task involved three movement modalities: imagined (IM), passive (PM), and congruent imagined and passive (IPM) finger flexion. We found that alpha (8-13 Hz) and beta (13-30 Hz) band desynchronization in the sensorimotor area was significantly stronger for PM and IPM tasks compared to IM, with no significant difference between PM and IPM. Using machine learning models, we achieved high accuracy in classifying exoskeleton-assisted movements from the rest condition (IPM vs. REST: 0.80 ± 0.07; PM vs. REST: 0.72 ± 0.10), with the IPM modality returning the highest accuracy. However, distinguishing between IPM and PM yielded only 0.61 ± 0.09, significantly lower than the condition of intention detection without the exoskeleton (IM vs. REST: 0.73 ± 0.08). Our findings suggest that sensorimotor EEG activity can track proprioceptive feedback induced by the hand exoskeleton. While this feedback is pronounced and distinguishable, detecting motor intention during exoskeleton movement remains highly challenging. This highlights the need for advanced decoders and control strategies for the future development of continuous BCI-actuated hand exoskeletons.

The HTTP/1.1 protocol has long been a staple on the web, for both pages and apps. However, it has started to show its age, especially with regard to page load performance and the overhead it entails due to its use of multiple underlying connections. Its successor, the newly standardized HTTP/2, aims to improve the protocol's performance and reduce its overhead by (1) multiplexing multiple resources over a single TCP connection, (2) using advanced prioritization strategies, and by introducing new features such as (3) Server Push and (4) HPACK header compression. This work provides an in-depth overview of these four HTTP/2 performance aspects, discussing both synthetic and realistic experiments, to determine the gains HTTP/2 can provide in comparison to HTTP/1.1 in various settings. We find that the single multiplexed connection can actually become a significant performance bottleneck in poor network conditions with high packet loss and that HTTP/2 rarely improves much on HTTP/1.1, except in terms of reduced overhead. Prioritization strategies, Server Push, and HPACK compression are found to have a relatively limited impact on web performance, but together with other observed HTTP/2 performance problems, this could also be due to faulty current implementations, of which we have discovered various examples.

Web page performance is becoming increasingly important for end users but also more difficult to provide by web developers, in part because of the limitations of the legacy HTTP/1 protocol. The new HTTP/2 protocol was designed with performance in mind, but existing work comparing its improvements to HTTP/1 often shows contradictory results. It is unclear for developers how to profit from HTTP/2 and whether current HTTP/1 best practices such as resource concatenation, resource embedding, and hostname sharding should still be used. In this work, we introduce the Speeder framework, which uses established tools and software to easily and reproducibly test various setup permutations. We compare and discuss results over many parameters (e.g., network conditions, browsers, metrics), both from synthetic and realistic test cases. We find that in most non-extreme cases, HTTP/2 is on a par with HTTP/1 and that most HTTP/1 best practices are applicable to HTTP/2. We show that situations in which HTTP/2 currently underperforms are mainly caused by inefficiencies in implementations, not due to shortcomings in the protocol itself.

The N-Back working memory (WM) updating task is commonly used in cognitive neuroscience research. Many studies investigated behavioral responses, while only few studies have sought to identify its electrophysiological signatures. Nevertheless, it is still unclear what the role is of stimulus type and task structure in the outcomes. Here, we address this issue using nine different variations of an N-Back task. Results show differences in electrode clusters between words vs. pictures for both N2 and P3-ERP components, specifically in the fronto-central brain regions. In conclusion, our findings suggest that stimulus type contributes to differences in electrophysiological responses during N-Back task performance, which needs to be accounted for when interpreting findings across studies.

The popularity of algebraic effect handlers as a programming language feature for user-defined computational effects is steadily growing. Yet, even though efficient runtime representations have already been studied, most handler-based programs are still much slower than hand-written code. In this paper, we show that the performance gap can be drastically narrowed (in some cases even closed) by means of type-and-effect directed optimizing compilation. Our approach consists of two stages. Firstly, we combine elementary source-to-source transformations with judicious function specialization in order to aggressively reduce handler applications. Secondly, we show how to elaborate the source language into a handler-less target language in a way that incurs no overhead for pure computations. This work comes with a practical implementation: an optimizing compiler from Eff, an ML-style language with algebraic effect handlers, to OCaml. Experimental evaluation with this implementation demonstrates that in a number of benchmarks, our approach eliminates much of the overhead of handlers and yields competitive performance with hand-written OCaml code.

Brain-Computer Interfaces (BCIs) are hailed for bypassing defective neural pathways by translating brain activity directly into actions that convey the user's intent. How the kinematics of muscular activity relate to the motor- and somatosensory activity in the brain has been the focus of recent advancements. With such motor BCIs, amputees are able to gain control over a prosthesis and stroke patients to regain control over a paralyzed limb via electrical stimulation of their dysfunctional muscles or via an exoskeleton that supports the intended movements. The superior spatio-temporal resolution, bandwidth, and recording stability of electrocorticography (ECoG), a partially invasive brain recording technique, yields a new outlook on motor BCI applications. Despite some stunning successes in arm- and hand movement control from ECoG, the precise decoding of finger movements, which is essential for daily activities, is still lacking. A possible reason is that current decoders rely on conventional one- or two-way regression models, which might not adequately capture the intricate relation between neural activity and intended and unintended (such as coactivations) finger movements. The main objective of this PhD is to develop a robust, accurate, and quick-to-train decoder that predicts single- and coordinated finger trajectories from ECoG recordings. We used multiway decoders as they preserve the multilinear structure of the data while taking advantage of potentially hidden multilinear components. We demonstrated cutting-edge performance with the proposed decoders. As multiway models tend to be slow to train, which may become a significant obstacle for their clinical adoption, we also investigated whether the proposed multiway decoders could be used in a real-time setting. The findings support the relevance of the proposed multiway decoders for real-time ECoG-based finger activity, providing an outlook on achieving hand dexterity.

Determining how distinct brain regions are connected and communicate with each other will shed light on how behavior emerges. In EEG studies, interpreting connectivity measures can be problematic due to the high correlation between signals recorded from the scalp surface, a result of the volume conductance of the scalp and skin. Therefore, meaningful connectivity patterns can be measured only from the spatiotemporal distribution of localized cortical sources, generally referred to as source reconstruction. Still, spurious connectivity issues may persist in source-reconstructed EEG data, making it vital to choose an appropriate measure of connectivity. In this work, an information theoretical approach, which concerns model-free, probability-based methods such as mutual information, conditional mutual information, and interaction information, is taken. An information theoretical framework for Python is developed to operate on source-reconstructed activity. This framework is used to perform a connectivity analysis of a high-density source-reconstructed EEG dataset, which was constructed in an experiment regarding the semantic processing of abstract and concrete words.

Algebraic effects and handlers benefit from a custom type-&-effect system, a type system that also tracks which effects can happen in a program. Several such type-&-effect systems have been proposed in the literature, but all are unsatisfactory. Recently, Stephen Dolan (University of Cambridge, UK) presented a novel type system that combines subtyping and parametric polymorphism in a particularly attractive and elegant fashion. A cornerstone of his design is the algebraic properties that the subtyping relation should respect. In this work, a type-&-effect system is derived that extends Dolan's elegant type system with effect information. This type-&-effect system inherits Dolan's harmonious combination of subtyping (in our case induced by a lattice structure on the effect information) with parametric polymorphism and preserves all of its desirable properties (both low-level algebraic properties and high-level meta-theoretical properties like type soundness and the existence of principal types). This type-&-effect system has been implemented in the Eff programming language in order to provide a proof-of-concept.

Large data centers are storing and sending more and more data. In order to check whether the network traffic does not contain intrusions, an intrusion detection system is used. Such a system analyzes data from the network and gives an alert if it finds an intrusion. Since data centers have so much data traffic, it is difficult to process everything. That's where IP flows come into the picture. They are aggregated from packet data but do not contain any information about the payload data. This thesis explains which attacks can be detected and how IP flows can be used for intrusion detection. It would also be cost-efficient if an intrusion detection system could operate automatically and detect attacks with a high probability. For this, machine learning can be used. Machine learning is a type of Artificial Intelligence which allows programs to learn and find patterns within data. However, there are many different types of machine learning. This thesis gives an introduction to machine learning concepts and provides an overview of different machine learning algorithms such as Support Vector Machines and K-Nearest Neighbors. An explanation is given on how these algorithms can be used in an intrusion detection system. The algorithms are evaluated on different datasets which consist of both labeled training data and unlabeled real-world data. The evaluation is done using learning curves and F-scores. In the evaluation, it is found that supervised learning gives better and more detailed predictions compared to unsupervised learning. K-Nearest Neighbors gives the best results among the tested supervised learning algorithms. The results show that machine learning is a viable option to detect intrusions using IP flows. It can also be noted that using extra information such as the TCP flags is a useful addition and increases the performance significantly.